An Unsupervised Learning Approach for Multimodal Low Back Pain Stratification

Narasimharao Kowlagi, Eveliina Heikkala, Simo Saarakkala, Jaro Karppinen, Aleksei TiulpinSpine 2025 · DOI

TL;DR

This paper introduces a novel, unsupervised machine learning framework that integrates deep learning-derived MRI biomarkers with traditional questionnaires and demographic data to stratify individuals with low back pain (LBP). This multimodal approach identifies distinct patient subgroups, offering a more personalized and potentially effective strategy for LBP management compared to current methods.

What Problem Does This Paper Address?

Low back pain (LBP) is a complex condition with diverse underlying causes, making effective treatment challenging. Current approaches often rely on questionnaires like the STarT Back Tool (SBT) and Örebro Musculoskeletal Pain Screening Questionnaire (ÖMPSQ) short to stratify patients based on psychosocial factors. While valuable, these tools may not fully capture the multifactorial nature of LBP, which can also be influenced by physical degenerative changes. This study aims to bridge this gap by developing a more comprehensive stratification framework that integrates both physical (imaging biomarkers) and psychosocial data, hypothesizing that this multimodal approach will reveal clinically meaningful subgroups not discernible by single-modality methods.

Methods

The core of this study is a novel, unsupervised learning framework for patient stratification. The approach involves:

  • Data Integration: Combining demographic data (sex, BMI), lifestyle information (smoking), psychosocial data (from SBT and ÖMPSQ-short questionnaires), and advanced deep learning (DL)-derived imaging biomarkers from lumbar spine MRIs. These imaging biomarkers include Pfirrmann grades (for disc degeneration, DD) and facet tropism (FT) angles.
  • Deep Learning for Biomarker Extraction: Utilizing pre-existing DL models to extract quantitative imaging biomarkers from T2w lumbar spine MRIs.
  • Dimensionality Reduction: Employing Uniform Manifold Approximation and Projection (UMAP) to reduce the dimensionality of the integrated dataset while preserving essential data structures.
  • Unsupervised Clustering: Applying a robust K-Means clustering algorithm, performed separately for SBT and ÖMPSQ-short data, to identify distinct patient subgroups. A five-fold cross-validation approach was used to ensure robustness, generating a centroid matrix for final cluster assignment.
  • Cluster Characterization: Analyzing the within-cluster distributions of LBP frequency and bothersomeness, and using logistic regression to identify significant features associated with each cluster.
  • Predictive Modeling and Validation: Training predictive models (logistic regression) to assign participants to clusters and evaluating their performance using precision-recall and ROC curves.
  • Clinical Utility Assessment: Employing Decision Curve Analysis (DCA) to assess the clinical net benefit of the stratified approach compared to a "treat-all" strategy.

Datasets

The study utilized retrospective, cross-sectional data from the Northern Finland Birth Cohort 1966 (NFBC1966). This dataset includes:

  • Demographic information (sex, BMI).
  • Lifestyle habits (smoking status).
  • Psychosocial characteristics captured via the STarT Back Tool (SBT) and Örebro Musculoskeletal Pain Screening Questionnaire (ÖMPSQ) short.
  • T2-weighted (T2w) lumbar spine MRIs acquired when participants were 46 years old.

No new benchmark datasets were introduced; the study leverages existing, well-characterized cohort data.

Key Insights and Ideas

  • Multimodal Data is Key: The study strongly suggests that combining imaging biomarkers with psychosocial data is crucial for a comprehensive understanding of LBP and for effective patient stratification.
  • Distinct Subgroups Emerge: The unsupervised approach successfully identified three distinct subgroups, each with a unique profile characterized by varying degrees of physical degeneration and psychosocial risk.
  • Imaging Biomarkers Add Value: Deep learning-derived MRI biomarkers (DD and FT) are not just descriptive but actively contribute to differentiating LBP patient profiles, offering a more objective layer of information.
  • Clinical Utility Demonstrated: The stratified approach showed a superior net benefit compared to a "treat-all" strategy, indicating its potential to guide more personalized and effective treatment decisions.

Key Results

  • Three Clusters Identified: The K-Means clustering algorithm, optimized with UMAP, consistently identified three distinct subgroups for both SBT and ÖMPSQ-short analyses.
  • Cluster Profiles:
    • One subgroup was dominated by psychosocial characteristics.
  • Another was characterized by physical degenerative changes (DD).
  • A third subgroup represented a mix of both.
  • Predictive Model Performance: The predictive models for cluster assignment achieved high mean accuracies: 0.89 ± 0.01 for the SBT-based model and 0.87 ± 0.02 for the ÖMPSQ-short model. These models significantly outperformed random assignment.
  • Superior Net Benefit: Decision Curve Analysis demonstrated that the stratified approach consistently offered greater utility and a superior net benefit compared to a "treat-all" strategy across various threshold probabilities.

Limitations and Future Work

The authors acknowledge several limitations:

  • Cross-Sectional Design: The study's cross-sectional nature prevents causal inferences and the observation of longitudinal changes in LBP progression.
  • Limited Imaging Biomarkers: The study primarily focused on disc degeneration and facet tropism. Incorporating a broader range of imaging biomarkers (e.g., disc herniation, spinal stenosis, paraspinal muscles) could further enhance stratification.
  • Reliance on Self-Reported Data: The use of questionnaires introduces potential biases such as recall bias and social desirability bias.
  • Potential for Measurement Bias: Clinical examinations might also introduce measurement bias.
  • Need for External Validation: The findings require external validation in different cohorts to ensure generalizability.

Future work should focus on:

  • Longitudinal Studies: Investigating the progression of LBP and treatment outcomes over time within these identified subgroups.
  • Broader Biomarker Integration: Including a wider array of imaging and potentially other biological markers.
  • Tailored Interventions: Developing and testing specific intervention strategies for each identified cluster.
  • External Validation: Replicating the study in diverse populations to confirm the robustness and generalizability of the findings.

Summary updated

Abstract

A novel framework was developed that integrates multimodal data to identify distinct subgroups differentiated by their physical and psychosocial characteristics in a population-based cohort, demonstrating potential for advancing personalized care.
BibTeX
@inproceedings{kowlagi2025unsupervised,
  author = {Narasimharao Kowlagi and Eveliina Heikkala and Simo Saarakkala and Jaro Karppinen and Aleksei Tiulpin},
  title = {{An Unsupervised Learning Approach for Multimodal Low Back Pain Stratification}},
  booktitle = {Spine},
  year = {2025},
  doi = {10.1097/brs.0000000000005593},
}
Figure